Thread: Re: [HACKERS] ALTER TABLE .. ALTER COLUMN .. ERROR: attribute .. has wrong type

Alvaro Herrera <alvherre@2ndquadrant.com> writes:
> Tom Lane wrote:
>> [ pokes around... ]  The code I was thinking of is convert_tuples_by_name
>> in access/common/tupconvert.c.  There's a bit of an API mismatch in that
>> it wants to wrap the mapping array in a TupleConversionMap struct; but
>> maybe we could refactor tupconvert.c to offer a way to get just the map
>> array.

> Ah, nice gadget.  I think the attached patch should do.

Looks reasonable to me.

>>> I also modified the algorithm to use the relcache instead of walking the
>>> child's attribute list for each parent attribute (that was silly).

>> Hmm.  That might be better in a big-O sense but I doubt it's faster for
>> reasonable numbers of columns.

> Hm, I was thinking in unreasonable numbers of columns, keeping in mind
> that they can appear in arbitrary order in child tables.  Then again,
> that probably seldom occurs in real databases.  I suppose this could
> become an issue with table partitioning becoming more common, but I'm
> okay with deferring the optimization work.

It occurred to me that it'd be really easy to improve
convert_tuples_by_name so that, rather than having the inner loop
start from j = 0 every time, it starts from the attribute after the
last match (and loops around if needed, so that it still examines
every child attribute).  I think this would keep it at more-or-less
linear time for all but very contrived child tables.

Since your patch is touching that code I won't do anything about it
right now, but maybe later.
        regards, tom lane

Tom Lane wrote:
> Alvaro Herrera <alvherre@2ndquadrant.com> writes:
> > Tom Lane wrote:
> >> [ pokes around... ]  The code I was thinking of is convert_tuples_by_name
> >> in access/common/tupconvert.c.  There's a bit of an API mismatch in that
> >> it wants to wrap the mapping array in a TupleConversionMap struct; but
> >> maybe we could refactor tupconvert.c to offer a way to get just the map
> >> array.
> 
> > Ah, nice gadget.  I think the attached patch should do.
> 
> Looks reasonable to me.

Thanks for looking!  Pushed.

> > Hm, I was thinking in unreasonable numbers of columns, keeping in mind
> > that they can appear in arbitrary order in child tables.  Then again,
> > that probably seldom occurs in real databases.  I suppose this could
> > become an issue with table partitioning becoming more common, but I'm
> > okay with deferring the optimization work.
> 
> It occurred to me that it'd be really easy to improve
> convert_tuples_by_name so that, rather than having the inner loop
> start from j = 0 every time, it starts from the attribute after the
> last match (and loops around if needed, so that it still examines
> every child attribute).  I think this would keep it at more-or-less
> linear time for all but very contrived child tables.
> 
> Since your patch is touching that code I won't do anything about it
> right now, but maybe later.

Yeah, I had the same idea.  Looks fiddly but not terribly difficult, and
well localized.  Didn't we have a list of tasks for eager contributors?

-- 
Álvaro Herrera                https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services